Modification of Bayesian Optimization for Efficient Calibration of Simulation Models

Simulation models contain many parameters that must be adjusted (calibrated) in advance to reduce the error between simulations and experimental results. Bayesian optimization is often applied to minimize error after only a few simulations. However, Bayesian optimization uses only error information, ignoring information on other simulation results. In this paper, we improve Bayesian optimization by utilizing both and show that other simulation results effectively reduce the dimensionality of the parameter space. In an evaluation using actual semiconductor simulation results, the proposed method reduces the number of simulations by 50% compared with random search and conventional Bayesian optimization.